Estimating Highway Construction Production Rates during Design: Elements of a Useful Estimation Tool

Productivity factors occurring on “normal” and “abnormal” production days may be different (Abdul Majid and McCaffer 1998). Designers seek to estimate a normal production day, as an abnormal production day includes variables that cannot be controlled, such as rain heavy enough to prevent excavation work, a road accident, or unforeseen utility conflicts or soil conditions. Excusable delays occur only on abnormal days, and contractors are normally allowed to claim time extensions for these delays. As such, productivity factors on normal and abnormal days are separated, and designers base their estimates on a normal production day.

One of the most common scheduling practices during the design phase is to break down construction work into constituent operations and estimate each operation individually. There is no standard way of breaking down the operations; it depends on how detailed the designers wish to be. Breaking down each work operation into smaller pieces for estimating purposes may increase schedule reliability; however, limited knowledge of and control over operations may prevent designers from doing so. Designers may not have the luxury of time and information to estimate their schedule in such detail. Most designers break operations down into a size that is manageable enough to increase the schedule’s reliability without involving too many variables. Hancher et al. (1992) found that the common practice among designers was to break down their production rates into work items, each of which has its own productivity variables.

Schedule estimates are then based on critical work items on a project. CPM involves identifying the critical path, which consists of work items that take the longest overall time to complete on a project and thus normally extend over the project completion time. Activities that fall on the critical path are what designers need to schedule for. These activities normally include work items with the highest quantities or those that require a lot of lead time. A typical highway project includes more than 100 work items. Given the tight schedule in most projects, designers do not have the luxury of estimating every work item. They need to identify and concentrate only on the critical work items in their projects.

Highway construction time estimation methods adopted by different states, agencies, and districts can be significantly different from one another. The Texas DOT, for example, adopted the Construction Time Determination System (CTDS; Hancher et al. 1992), and the New Jersey DOT adopted the Capital Program Construction Scheduling Coding and Procedures for Designers and Contractors Manual. These scheduling methods are quite different. For the New Jersey DOT, the production rate for a cast-in-place (CIP) retaining wall for Type 1 bridge construction is 20 days/30 meters, given the following assumptions: (1) 8-hour working day per crew, (2) 50% added for one bridge and 25% added for two, (3) separate estimates for bridges built in different stages, (4) additional 30 days required for bridges built over water, and (5) inclusion of concrete curing time. In contrast, the Texas CTDS production rate for retaining wall construction is 9.3 to $18.5{\mathrm{m}}^2$ per day, and soil condition is the only adjustment needed. Thus, the two methods differ in units (area versus length), productivity factors, and design issues. In both systems, however, designers’ experiences are critical in their decisions regarding the ideal production rate based on the range given.

The Transportation Research Board conducted a series of studies between 1981 and 1995 to investigate and develop systems to improve the reliability of contract time estimation of highway construction projects (National Cooperative Highway Research Program 1981; Herbsman and Ellis 1995). In addition, Hancher et al. (1992) found in a survey of employees in 36 DOTs that 44% of the respondents relied on personal experience to estimate production rates, 30% used standard production rates that were usually provided by the DOTs, and 22% used production rates from historical records of previously completed projects from other sources. Data used to develop these systems were normally collected either from survey returns or documentation from projects. Some DOTs and several contractors had simplified their production rate estimates into three values—mean, optimistic, and pessimistic values for each work item—and designers determined a realistic rate from these values based on their experience.

In production scheduling, productivity factors are used to establish relationships between fluctuating production rates and production work items. Designers must identify relevant productivity variables they can rely on to improve their estimates.

Differing geological locations, site conditions, and other location variables affect work productivity. Constructing a drilled shaft on dry soil may be faster than constructing one nearer a riverbank, assuming the same crew is used in both operations. Constructing a road on a mountain may be much slower than constructing one on flat land. A sudden change in the weather can disrupt or delay construction operation. Construction work has to stop during snowstorms in the warmer regions (e.g., Florida and Texas) but can continue in the colder regions (e.g., Maine and Washington). Rural highway projects face fewer traffic problems than highway projects in metropolitan areas, which face frequent traffic congestion, strict environmental regulations, risky traffic safety, and right-of-way issues that reduce construction productivity. Rural highway construction, however, may face difficulties in acquiring skilled workers and highly productive equipment, driving down productivity (Koehn and Ahmed 2001). Thus, the location of the project is not a dependable productivity factor; in a work zone located in a metropolitan area that has ample work space, highway congestion may affect only material delivery and may have minimal impact on work operations.

Various productivity factors have different impacts on materials and components. For example, off-site fabricated precast concrete has less productivity variability during installation but a higher chance of being delayed during transport, whereas CIP concrete has a higher chance of being delayed by poor weather conditions. Other productivity factors include size of equipment (Bhurisith and Touran 2002; Sawhney and Mund 2002), controllability of soil condition (Allouche et al. 2001), weather conditions, and losses from learning (Thomas et al. 1999; El-Rayes & Moselhi 2001). For example, it makes sense to link the productivity of foundation, pipeline, and retaining wall construction to soil conditions. Alternatively, the learning curve and poor weather conditions have a greater effect on highly repetitive work items such as pavement, multiple drilled shafts, and hot-mix asphalt pour. Thomas et al. (1999) found that changing weather, erroneous work, poorly coordinated material delivery, and frequent equipment relocation were more disruptive to long and continuous production processes than to less repetitive work items. Stoppage to a process slows down work momentum and leads to productivity losses of other work, which Thomas et al. (1999) described as the “ripple effect.” However, because a highway designer cannot accurately predict the chances of having disruptions during construction, such disruptions are normally excluded from the schedule at the design phase.

Design can significantly affect work productivity (Poh and Chen 1998). Constructability has been shown to increase site productivity, and site congestion can be avoided with designs that use smaller equipment.

Site condition would normally be considered a productivity factor. Construction on mountainous regions, in tight work spaces, in extreme cold and heat, on rough terrain, on congested work zones, and in close proximity to adjacent structures would normally slow down workers’ productivity (Koehn and Brown 1985).

Wideman (1994) showed that the productivity of workers varied during different phases of construction. His study found that workers’ productivity was slow during the early phase of construction and slowly sped up as construction progressed. Productivity continued to rise and plateau between 25% and 75% of project completion, and then fell until project completion. He attributed the initial growth to workers’ learning effect and the fall nearing project completion to reduced work amount. Thus, worker productivity should not be treated as uniform throughout the construction phase.

Most research has found that work productivity is affected by more than one productivity factor. For example, our literature review indicated that foundation construction is affected by many factors, including soil type, drill type (size, type, and construction method), angle of swing, methods of spoil soil removal, pile axis adjustment, depth and size of holes, equipment power, operator efficiency, weather conditions, spoil soil removal and space availability, rebar cage installation procedures, concrete pouring methods, machine availability, job and management conditions, drilling time activity, other time activities, change orders, and weather (Zayed and Halpin 2004; Hanna et al. 2002; El-Rayes and Moselhi 2001). Furthermore, Thomas et al. (1989) emphasized that productivity factors should be divided into within-project, between-project, and regional drivers. Such divisions allow designers to better identify and allocate significant productivity drivers for different work zone scenarios.

Developing reliable productivity data documentation and analysis processes is critical to the development of a dependable scheduling tool. Data accuracy and correct analysis are the two critical elements of any successful and reliable information system. Collected productivity data should support the following functions: (1) identify significant productivity factors, (2) examine and establish the relationships between factors and productivity, (3) develop production rate models for scheduling, and (4) study productivity improvement methods. Efficiency of data collection is also critical, though there is a need to balance efficiency and reliability. Survey forms and existing productivity databases are the two most common data collection methods. Hancher et al.’s (1992) survey of experienced designers and site personnel in 36 DOTs indicated that many contractors kept and developed their own production factor and rate information. They gathered most of their productivity information from daily log books, payment and schedule documentation, record books, and information stored on computers. Some contractors required their staffs to input information systematically and kept comprehensive records of all their projects. No procedures or standards have been developed for such recording processes; thus, the reliability of the information may vary greatly across organizations. Like any research using the survey approach, the reliability of Hancher et al.’s findings cannot be verified, and most of the data are based on personal opinion. Although contractors may collect huge amounts of data from existing records, we question the reliability and usefulness of these data if they are not being recorded for production rate estimation purposes.

Regression analysis, factor analysis models, Monte Carlo simulation, fuzzy logic, schedule algorithms, and neural networks have been applied to scheduling (AbouRizk and Wales 1997; Adeli and Karim 1997; Ben-Haim and Laufer 1998; Jiang and Shi 2005; Lee 2005; Lee and Arditi 2006). Several attempts have been made to integrate some of these techniques with standard scheduling software, such as Primavera, Microsoft Project, and SureTrak. Alternatively, many designers continue to develop project schedules using their own experience. Advanced scheduling techniques may improve a schedule estimate’s accuracy, but many departments of transportation lack the required infrastructure and information to use these techniques. In addition, scheduling tools developed in the past for specific purposes became outdated very quickly. Texas DOT’s CTDS (Hancher et al. 1992), for example, became obsolete simply because there was no way to update the information in the system. Thus, any new scheduling tool should support popular scheduling software, such as Primavera and Microsoft Project.

Lessons learned from the past have shown that schedulers tend to resist a complicated information technology system and prefer a flexible system. Because most designers use the critical path and Gantt chart methods to schedule their projects, Primavera and Microsoft Projects are the most frequently used scheduling programs. Therefore, production rate estimation tools should not deviate too much from these techniques and software. Production rate information that is easily integrated with the software and techniques will allow designers to spend less time learning new techniques and software and more time improving the reliability of their estimates. Consequently, any production rate system should be either completely independent from or well integrated into existing software and estimation techniques.

Researchers have successfully applied many techniques to improve the reliability of production rate estimates. Some of these techniques include regression analysis and models, factor analysis models, Monte Carlo simulation, and neural networks. Other, more simplified methods include summing up collected production rates and averaging them into mean, optimistic, and pessimistic values (Hancher et al. 1992). The techniques an organization adopts depend heavily on the types of obtainable and available information within the organization, the predictability of the productivity factors, and the details and accuracy requirements of the production rate estimates. Neural network application to estimation requires an extensive amount of information to develop the network to the point where the network would self-learn and correct itself.

Lu and AbouRizk (2000) supplemented PERT (Program Evaluation and Review Technique) with statistical and simulation techniques to develop a six-value estimation technique: minimum duration (optimistic), maximum duration (pessimistic), mean, standard deviation, confidence interval, and probability. This technique allows estimators to better understand the risks and confidence of their estimates and eliminates the need to guess. Regression analysis is one of the most common methods applied to examine and quantify the relationships between productivity factors and rates (Koehn and Brown 1985; Sanders and Thomas 1993; Hanna et al. 2002). Regression analysis is useful to illustrate the continuous relationships between numeric productivity factors and rates. However, it cannot be used to develop relationships between nonnumeric (categorical) productivity factors and rates. For such factors, Hancher et al.’s (1992) method is the most common approach; the contractor determines a productivity rate under normal conditions (normally, a mean or median) and two rates for extreme conditions (normally, optimistic and pessimistic). Other statistical techniques are available to handle nonnumeric factors (categorical factors), such as box plots and longitudinal data analysis techniques (commonly used in social science research). The more advanced techniques require special treatment of data, like creating artificial variables and splitting data into different components.

Peña-Mora and Li (2001) proposed using the GERT (Graphical Evaluation and Review Technique) diagramming scheme to calculate the probability of project duration by measuring the variability of different branches and loops of each construction activity, their relationships, and overlaps with other nonrelevant activities. They claimed that such a scheme could better control or even eliminate variability within a schedule. Park and Peña-Mora (2004) proposed using reliability control to refine activity buffers and applied simulation to measure and reduce buffer variability between activities. However, designers may not have sufficient control and may lack the ability to predict buffer variability between activities during design, and they cannot feasibly apply both techniques at the same time. Also, these techniques are applicable only for measuring relationships between activities and cannot be used to improve the reliability of production rates.

Many researchers have applied dynamic and stochastic approaches and developed simulation models that integrate the construction process network using activities at the project level to develop production rate models that can improve the reliability of duration variability estimates of and between activities. AbouRizk and Wales (1997) illustrated that such a project simulation model should consist of three components: (1) a project network that maintains schedule logic, (2) a stochastic and random particles model that generates uncertain factors, and (3) a productivity model that relates uncertainty of productivity factors to generated project conditions. Each activity could be simulated individually and combined at a later stage (discrete-event continuous simulation). Discrete events could be combined for simultaneous and continuous simulation. Their models relied on historical data such as weather data from meteorological agencies. Construction processes are broken down into individual activities based on their relationships with various productivity factors. The effect of a productivity factor on an individual activity is measured by the simulation process and is later combined into a schedule.

Dzeng and Tommelein (1997) suggested breaking down construction projects into “cases” and automating the duration estimating process for each case, and they found that their schedule estimates were more accurate. The proposed application of neural networks to construction scheduling (Adeli and Karim 1997), although seeming to help improve the reliability of construction scheduling through continuous knowledge learning, has limited application at the design stage, especially on improving the reliability of production rate estimates. Adeli and Karim (1997) stated that neural dynamic models require breakdown of work into tasks, crews, and segments, and logics and constraints between repetitive tasks have to be developed while each task is simulated individually. Excessive amounts of data are required to ascertain the usefulness of neural dynamic models since the accuracy of the final estimates heavily depends on inputs to the system. Such requirements of massive and continuous information inputs make the neural network inefficient and impractical to be applied at the design phase.

Bonnal et al. (2004) believed that fuzzy logic had become sufficiently mature to be applied to project scheduling in real life. They claimed that fuzzy logic could eliminate calculation imprecision, narrow the probability of estimates, and improve the plausibility and credibility of calculated values. In most production rate estimates at the design phase, uncertainties during production cannot be predicted, and thus designers cannot include them in the schedule. Fuzzy and probabilistic statistical techniques can be extremely complicated. The need to create new variables (including dummy ones) to simulate uncertainties of individual activities and between them (Vanhoucke 2006; Fan and Tserng 2006) requires designers to look at different ways of treating their data. In real life, these variables do not exist, so it may be difficult for designers to understand and appreciate their meanings and purposes. These techniques also require extensive algorithms to estimate other resource constraints (Jiang and Shi 2005). Algorithms are needed to clarify relationships between activities and allow schedulers to better control variability between activities. Refinement and reliability of the schedules could be enhanced with algorithms.

Construction schedules are dynamic and thus need to be updated frequently to reflect changes during construction. El-Shahhat et al. (1995) found that between 29% and 67% of all construction errors occurred at the design phase, while only 12% to 59% of errors occurred at the construction phase. Errors committed at the design phase may be magnified at the construction phase and cause problems during construction operations. Thus, designers cannot rely too much on their personal experience.

A survey conducted by the Texas Department of Transportation found that scheduling practices varied across districts and regions (O’Connor et al. 2005). Most of the practices were driven by the needs of the districts, the top management, and the demands of the engineers in those districts. Some districts even developed their own system to handle their unique needs and situations. Each district had authority over its own scheduling methods, and employees were given the option to adopt anything that suited them. Scheduling software used in the Texas DOT included in-house systems like Primavera, Microsoft Project, and SureTrak. The type of software used depended heavily on the preferences of the employees. Several districts did not use the Texas DOT’s standard system (i.e., CTDS), as they found that it was inflexible and that the information was not accurate. They preferred the flexibility of using any software they chose. Many districts preferred user-friendly and flexible systems that could easily integrate into Primavera, Microsoft Project, and SureTrak. Indeed, most highlighted the tendencies to use CPM and/or Gantt charts for scheduling. They preferred to carry out production rate calculation manually rather than using the software.

Many respondents suggested they would prefer a system that would integrate with Microsoft Excel. They highlighted that because some of the projects were small, manual calculation helped speed up the estimation process. The feedback also highlighted that accurate production rate information was needed that reflected and represented the situations and conditions in different districts and regions, like calendar day and working day contracts, rural areas, soil conditions, and regional profiles. Many designers highlighted that the units adopted should be in line with the pay units, as this would reduce their work to convert the units between schedule and payment.

We selected eight Texas DOT designers for interview with regard to their needs for a better scheduling tool. Their feedback included the following seven points:

We identified the following six key components of a production rate estimation system:Integrating these components into current schedule estimation practices is critical; they are consistent with existing good schedule estimating practices. In addition, several useful schedule estimation techniques are available to estimate and analyze production rates, such as statistical analysis (normally regression), neural networking, Monte Carlo simulation, and linear programming.

Most designers surveyed for this research conducted their scheduling using the CPM and/or Gantt charts. They often relied on Primavera, MS Project, or SureTrak, so any scheduling tools should be either integrated into or separate from this software. The designers wanted to retain the functions of CPM and Gantt charts but needed more reliable and useful information on production rates and production variability to develop better estimates. Thus, they were looking for reliable and useful production rate and variability estimation tools that could easily be integrated into or applied to existing CPM or Gantt charts. In addition, because it is impossible to collect sufficient data to establish good statistical models for every factor, designers need some freedom to use their experience to adjust the production rates provided by these tools in order to estimate more realistic rates. Relying solely on experience can make production rates less reliable and unrealistic, as research has shown that human memory distorts information easily; likewise, scheduling tools alone cannot replace experience. Combining experience with reliable scheduling tools is the solution to more reliable construction project scheduling.

Scheduling tools should apply only the less complicated statistical and simulation techniques. It is impossible to collect sufficient and reliable information to comply with the requirements of the more advanced techniques. The designers found that many values generated by these techniques did not mean anything to them and noted that schedules do not need to be accurate, just realistic and reliable. The more advanced statistical and simulation techniques help improve the accuracy of models; however, most project schedules do not require this kind of accuracy.

It is also important to ensure that data are collected efficiently and reliably. The sources of data are critical, and sources need to be properly selected and checked before the data are used to develop scheduling tools. Although breaking down work processes into greater detail can help improve the accuracy of estimates, it is impractical to expect designers to do this. Designers do not have control over the actual work processes, and scheduling is one of the many duties they have to do. Separating the production rates by work items is a more efficient way for designers to break down work processes. In addition, factors have to be properly described, meaningful to the designers, and foreseeable at the design stage. The literature reviews also found that productivity estimation tools may present the rates in many ways. The designers preferred to have a tool that presents flexible values or even ranges of values that they could choose from and would allow them to adjust according to their own logic. In short, they demanded controllability, applicability, and flexibility in the estimation tool.

Finally, productivity factors should be categorized appropriately and integrated when several factors have to be considered. Historical information without proper appraisal may not reflect the actual conditions when they were collected. Historical information on pipeline production, for example, can document rates as low as 10 meters per day or as high as 400 meters per day. Users of such information should have access to reasons, factors, or models to use in deciding whether their rates should be 10 or 400. Relying on personal experience to decide a rate will naturally incorporate some psychological effects.

Literature reviews, surveys, and interviews have highlighted important factors to be considered in designing scheduling tools. The results also highlighted several important requirements by practitioners that past researchers failed to appreciate. The industry needs a practical tool to support designers and integrate their experiences, and this research can be used to inform future work to develop such a tool.